U-PolyConformer: Spatiotemporal machine learning for microstructure engineering

• A novel spatiotemporal CNN-Transformer machine learning framework is proposed. • The first framework that captures full-field stress and strain evolution under arbitrary loading conditions. • Exhibits strong generalization; accurately capturing loading conditions, microstructures, and hardening outside the training distributions. • A novel microstructure optimization framework is employed to enhance formability by delaying localization. Accelerating the prediction of mechanical behaviour in heterogenous materials is critical for large-scale microstructure optimization and realizing functionally optimized materials. While existing machine learning approaches have demonstrated an ability to accelerate predictions for the full-field mechanical response of a wide range of heterogenous microstructures, they have been largely limited to monotonic loading conditions. This paper introduces U-PolyConformer, a spatiotemporal machine learning framework that combines U-Net convolutional neural networks with transformer layers, capable of capturing the full-field stress and strain evolution under monotonic and random walk loading conditions. Trained on a large dataset of crystal plasticity finite element method (CPFEM) simulations with FCC polycrystals, the model accurately captures complex phenomena, including strain localization and stress unloading. The U-PolyConformer achieves a 7,900x speed-up over the ground-truth CPFEM simulations while producing high-fidelity results in both interpolative and extrapolative regimes. Comprehensive evaluations demonstrate the U-PolyConformer’s capacity to generalize outside the training distribution to novel microstructures, loading conditions, and strain hardening behaviours. To highlight the model’s potential as a surrogate for accelerating computational materials engineering workflows, a microstructure optimization framework based on static recrystallization is introduced and used to delay the onset of localization. This framework is successfully used to identify the grains which initiate the onset of localization, illustrating how the proposed model and optimization framework may be used for identifying and exploring property-performance relationships.